Plenty of AI sessions cover what agents can build. This one covers the other side: what happens when an agent takes the wrong action, and the controls that prevent it.

In this session, we’ll cover:

• The common ways agents fail, and what each one costs

• Where to put a human approval step, and where it just adds friction

• Limiting what an agent can access and reviewing what it did

• How to decide which work stays with a person

If you’ve held off on agents because you weren’t sure you could keep them in check, this session is built for you.

Tuesday, August 25, 2026 · 12:30 PM ET / 11:30 AM CT / 9:30 AM PT · Free live webinar with Josh Sullivan, COO at Kiingo AI.
Register for the August 25 webinar →

• Anthropic’s annualized revenue run rate passed $65 billion at the end of July, up from $47 billion in May and about $9 billion at the end of last year. Preliminary second-quarter revenue hit $11.5 billion against $787 million a year earlier. The company has filed confidentially and could go public as soon as this fall. (Bloomberg, CNBC, TechCrunch)

• Nvidia disclosed it will guarantee up to $105 billion of debt financing for OpenAI’s new Ohio data center, covering an initial 4.25 gigawatts with an option for 3.75 more. The same day, a Wall Street Journal review found roughly $3 trillion in AI lease and purchase commitments across nine large tech companies, against about $600 billion of reported capital spending. (CNBC, Axios, WSJ)

• Stripe is paying $7.5 billion for OpenRouter, a company whose entire product is switching between AI models on a customer’s behalf. A business connects to it once, and each request goes to whichever of 400-plus models handles that job best or cheapest. The price is roughly five times what OpenRouter was worth in May. (CNBC, New York Times)

• AI-written text is starting to carry an invisible label. Driven by EU AI Act transparency duties in force since August 2, Anthropic will watermark Claude’s output using Google DeepMind’s SynthID and release a detection API, with other labs expected to follow. Google separately let users remove the visible watermark on Gemini images while the invisible signal stays attached. (Anthropic, TechCrunch, TechCrunch)

• Google won a bankruptcy auction for Spirit Airlines’ internal business data, paying $10 million for roughly 100 million emails, hundreds of millions of Teams messages, payroll records back to 1986, and pricing and operations data, for product development and AI training. Customer data on 97.5 million passengers was excluded, and Google must scrub personal information before receiving the files. (Axios, CNN, Skift)

• Microsoft patched a critical Copilot flaw on August 18 that let a single click on a malicious link pull data from a user’s connected accounts; Varonis reported it in December 2025. The same week, Microsoft merged its consumer and business Copilot apps, retiring Deep Research in favor of Researcher, which sits behind a $19.99-a-month Microsoft 365 Premium subscription. (Computerworld, BleepingComputer, TechCrunch)

• Data retention turned into a sales weapon. OpenAI is previewing Private Safety Processing, abuse monitoring that retains none of the customer’s data, targeting enterprises unhappy that Anthropic keeps prompts for 30 days on its most capable models; Bloomberg reports Anthropic plans to change that. OpenAI also shipped an Apple Messages plugin that reads message threads and drafts replies. (OpenAI, The Register, Engadget)

• IBM is building a dedicated OpenAI practice inside IBM Consulting, certifying tens of thousands of consultants, mostly retrained existing staff, and putting GPT-5.6, Codex, and ChatGPT Work into its consulting delivery platform. Joint offerings target financial services, telecom, government, and retail. IBM announced a near-identical alliance with Anthropic less than a year ago. (IBM, TechCrunch)

• Two models built for long-running agent work shipped a day apart, both with pricing that moves. SpaceXAI’s Grok 4.6 offers a 500,000-token context window at $2 and $6 per million tokens, doubling past 200,000 tokens. Google’s Gemini 3.7 Flash launched at an introductory $0.75 and $3.75, half its predecessor’s launch price, reverting to $1.50 and $7.50 on January 1. (SpaceXAI, SiliconANGLE, VentureBeat)

• Thrive Holdings raised more than $2 billion at a $12 billion valuation to keep buying ordinary service businesses, accounting firms and IT providers among them, and rebuilding their workflows around AI. It now operates more than 70 companies, and OpenAI, which took a stake in December, sends employees in to speed up adoption. (TechCrunch)

• Generalist AI published GEN-1.5, a robot model it says picks up a new physical task from a single demonstration with no retraining, averaging 59% success across ten tasks, rising to 83% after a brief tune-up on five minutes of data. The company says plainly the tasks are simple and short-horizon, and no outside lab has reproduced the results. (Generalist AI)

• Anthropic ran an autonomous protein-design campaign in which Claude designed binders for 15 targets and produced working ones for 14, with 22.6% to 35.1% of designs binding where 10% to 15% is typical. Adaptyv Bio and Twist Bioscience tested the molecules independently. Claude failed outright on one target, and the work ran on restricted research versions. (Anthropic)

• Google Cloud surveyed 2,403 executives and found 84% report increasing financial returns from AI. The 26% whose returns are accelerating year over year stand out on three things: extremely clear ownership of AI decisions, AI embedded in core business processes, and required training programs. (Google Cloud, July 27, 2026)

• The New York Fed’s regional business surveys report the share of service firms using AI rose to 40% from 25% a year earlier, and manufacturers to 26% from 16%. Firms report very few AI-driven layoffs and say they intend to retrain existing workers rather than replace them. (Liberty Street Economics, August 5, 2026)

• ServiceNow and ThoughtLab surveyed 4,500 senior leaders across 19 countries on how far AI has gotten inside their organizations. In Asia-Pacific, the 23% that reached the top tier of AI maturity report six times the productivity of everyone else. (ServiceNow Enterprise AI Maturity Index 2026, July 30, 2026)

You ask for a board update, a proposal, a client recap. Ninety seconds later a complete, confident, well-written draft arrives, and something about it is off. The sentences are fine. What’s wrong is the order of the argument, what got three paragraphs versus one line, and the section that should not be in there at all. So you rewrite the whole thing, and the assistant saves you very little. The fix costs about a minute: make it show you the structure before it writes a word of it.

1. Ask for the skeleton on its own. Give the assistant the task, the audience, and whatever material you normally would. Then add this:

“Before you write anything, give me the structure only: the sections in order, one line on what each will contain and roughly how long it will be, and the two or three points you plan to make the argument turn on. Mark anything you are unsure belongs. Do not write the draft yet.”

2. Correct the shape. Reorder it. Cut the section that exists because it usually exists. Tell it which part deserves the most room and which gets two sentences. This is the step where your judgment actually goes into the work, and reacting to five lines takes a minute.

3. Release the draft against the structure you approved. “Write it following the structure above exactly. If you deviate anywhere, note where and why at the end. If a section needs information I have not given you, say so in one line under that heading instead of filling it in.”

4. Keep the shapes that work. When one produces something you actually send, paste it back and ask for it as a reusable outline for future updates of that kind. Next time you start from a structure you already agreed with.

The draft is the cheap part. The structure is the part worth your attention, and correcting it while it is still five lines long is the whole trick.

Most of us have some kind of AI setup for email by now. A project, a saved prompt, something that helps get the wording right before a message goes out. What I keep coming back to lately is how much that same habit is worth on the small stuff: Google Chat, Teams, Slack, the messages nobody thinks of as writing.

Instant messaging carries a built-in sense of urgency. Something lands, and the format itself suggests you should already be typing. That pressure is worth checking. Running a message through AI first is a pass, the same way you would run a pass on anything else you send. A couple of minutes is usually the whole difference between a reply I fired off and one I actually meant.

I use it most when the message matters and I want to be absolutely sure. Something with a client in the thread. Something where tone will get read into it. Something I would rather not have to clarify twice. The chat is still real time. I just stopped treating real time as immediately.

The companies whose AI returns keep climbing share a short list of habits: someone owns the decisions, the tools sit inside the work rather than beside it, and training is expected. Every one of those is a choice you can start making this quarter.

Kiingo makes your company AI native: a sequenced plan for what to hand over next, hands-on training for the people who will use it, and clear ownership of every step along the way.

Let’s talk about where to start →